Most people who use Claude every single day have no idea how it actually works. And honestly, that’s not their fault. Nobody breaks down the mechanics in plain language, so you’re left with the surface stuff: it’s an AI assistant, it uses transformers, it can code. The real reasons behind its behavior stay hidden.
I came across a fantastic post from an AI professional who mapped out 30 facts about what’s happening beneath that surface. The creator built this out because founders in their network kept asking the same questions over and over. I was hooked the moment I started reading, because it explains so much of the weird behavior you’ve probably noticed.
The core idea from the author is simple: when you don’t understand the engine, you drive the car wrong.
The confusion this clears up
The original poster points out the exact frustrations most of us hit:
- You get confused when Claude forgets things mid-conversation
- You wonder why longer prompts sometimes drop in quality
- You can’t figure out why it refuses some requests but not others
Once you see the mechanics, every one of those quirks suddenly makes sense.
What’s happening under the hood
Here’s how the expert breaks down the basics. Claude is a large language model, just like ChatGPT and Gemini. The core transformer uses attention to weigh relationships between tokens all at once, not one at a time. It reads the entire input in parallel, not left to right like a human scanning a sentence.
After that attention step, Claude outputs a probability distribution and picks the most fitting next token. That loop runs token by token until the full response is done. Clean and simple when you see it laid out.
Where Anthropic does things differently
This is the part that really grabbed me. The creator highlights the design choices that make Claude its own thing:
- Constitutional AI (CAI): training is guided by a written set of principles, not just human preference scores
- Supervised phase: Claude self-critiques its own responses against that constitution
- RL phase: an AI model scores constitutional compliance to train a preference model
- Safety guardrails: baked into training itself, not bolted on as post-generation filters
- Operator authority: system prompts carry higher authority than user messages, within Anthropic’s hard limits
- Reasoning tokens: internal reasoning is stripped from context after each turn to preserve usable budget
- Extended thinking: a hidden scratchpad lets Claude reason step by step before answering
- No persistent memory: every conversation starts fresh
- Context rot: larger context helps with complexity but can degrade accuracy over time
- Production tricks: quantization, KV-caching, and dynamic batching cut latency and cost
Why this matters for how you build
The line that stuck with me most came straight from the author: model quality comes from architecture, data, and alignment working together. Not parameter count alone. That reframes how you think about prompting, tool use, and what you should realistically expect from the model.
Understanding the engine changes how you build with it.
My take and a strategic recommendation
Here’s why I think this is worth your time. Most teams are building on tools they only half understand, and that gap quietly costs them quality and trust. If you’re shipping anything with AI, treat this layer as required reading, not optional trivia.
My advice: pick two or three of these facts that map to a problem you’re hitting right now. Maybe it’s context rot on long prompts, or the fresh-memory reset between chats. Adjust your workflow around the actual mechanics instead of guessing. Then share the list with whoever on your team is using Claude without really getting it.
The creator packed way more nuance into the full breakdown than I can fit here. Head over to the original LinkedIn post to read all 30 facts and see which ones surprise you most.